Large language model agents increasingly support literature analysis, research ideation, experimen-tation, and scientific writing. Their collaboration depends on how source findings become research directions, how selected directions become experimental commitments, and how returned observations support manuscript claims. We introduce ResearchOS, an open-source multi-agent research system organized through research compilation, which gives these transitions explicit research artifacts and acceptance conditions. Source-linked reading and cross-paper synthesis establish a reusable research space. Within this space, a multi-agent evolutionary search develops structured study candidates using independent assessment, diagnosis-guided mutation, compatibility-aware crossover, and quality-diverse selection. Researcher-selected candidates undergo contribution review and study formalization before the approved plan is compiled into a contract for external execution. Returned records are checked against the study and connected to manuscript claims through evidence alignment and audit. An artifact-first runtime retains provenance, versions, researcher decisions, and recoverable state across the process. We present the architecture, evolutionary mechanism, execution interfaces, and representative system traces, illustrating how ResearchOS connects scientific exploration to executable investigations and evidence-supported reporting.